This paper provides asymptotic bias and variance analysis for MIMO
system estimates obtained by using generalizations of FIR model
structures and least squares techniques. The generalizations are such
that prior approximate knowledge of the system poles may be incorporated.
The obtained variance expressions provide extensions to well known
results that have previously been derived only for FIR model
structures. Namely, the asymptotic covariance of the transfer
matrix estimate is shown to be proportional not only to the
noise-to-signal ratio, but also to a frequency dependent term that
depends on the basis functions chosen. By examining a similar
expression for the bias error it is shown that it is not possible to
minimise the bias error at a particular frequency without increasing
the variance error, and vice-versa.